Visual saliency based abrupt motion object tracking

Aiming at solving the tracking problems under the circumstances of abrupt motion, a particle filter tracker is proposed based on visual saliency model. This tracker detects object from the salient regions in the saliency map by the way of winner-take-all and inhibition-of-return. The detecting result is taken as a global proposal distribution and then particles are sampled from it; therefore, the global state space can be searched in order to avoid suffering from the local minimum problem. Moreover, in order to increase the saliency of the object region in the saliency map, the bottom-up and top-down computational models are combined together. Then the weights of the feature maps are calculated according to the target template and saliency maps are fused adaptively. Compared with several other tracking algorithms, the experimental results of the proposed tracking method are more robust in dealing with various types of abrupt motion scenarios.